Skip to main content
QM pricing follows a fail-closed authority order:
  1. A finite positive QM cost is observed.
  2. Zero or negative QM cost is never observed money.
  3. A non-positive row is estimated only when its requested provider/model is explicit, deterministic, and present in Xerg’s date-aware catalog.
  4. Otherwise the observation is unpriced.
openrouter/auto is always unpriced unless QM eventually records the actual routed response model. Xerg does not guess a route. The negative placeholder values currently seen in QM are diagnostic only: tokens and cache buckets remain, but the values are excluded from known spend, savings, and identified-waste dollars. Because the current QM contract cannot distinguish a legitimate free-model zero from missing cost, Xerg says cost unavailable, not free or actual cost $0.

​
Certification cohorts

Release acceptance uses two separate staging workloads:
  • An explicit deterministic model with at least 95% priced token share, at least 95% priced underlying-request share, nonzero known spend, exact token-bucket reconciliation where QM exposes the authoritative buckets, and no negative value classified as observed.
  • An openrouter/auto workload with no placeholder spend priced, all token/cache evidence retained, and monetary thresholds exiting 5.
Setup never changes the production QM organization’s default model. A staging canary may select an explicit model only inside its private scope and only when a serviceable provider credential is already available through QM’s secret manager. Xerg runtime economics are not provider-invoice reconciliation. Use pricingCoverage in JSON output and the report notes to decide whether a monetary comparison is defensible.